In the following sections, an overview of ML and DL algorithms from an intelligent sensing perspective is presented. The aim is to highlight learning algorithms that are widely used in many real-time applications. Furthermore, parameters affecting the performance of intelligent sensing are also discussed. 

A machine that is able to make decisions on its own is said to possess AI. There is a broad spectrum of applications for AI, ranging from machine learning to robotics. By combining the current advancements in machine and deep learning, huge amounts of data from various sources are analyzed by utilizing AI to identify patterns and make intelligent predictions [51]. However, recent advances in artificial intelligence systems and robotics still need more research to solve complex problems. The tremendous growth in AI has ushered in a wave of applications using sensors. As a result, the demand for intelligent sensing increases in the market. Using sensor signals, the analysis of sensor data based on AI provides robust predictions and classifications. Hence, intelligent sensing will be the bright future of AI, where human behavior and emotions can be recognized by AI machines. 

The tremendous growth of ML-based approaches has expanded the research area of intelligent sensing. Generally, ML can be considered to be a subset of AI which handles complexities to solve a specific task. In this subsection, a brief overview of existing ML algorithms that improve the functioning of sensing systems is presented together with their advantages and disadvantages. Various scenarios portraying how machine learning methods are applied in intelligent sensing is depicted in Figure 2. ML algorithms are divided into supervised, semi-supervised, unsupervised, and reinforcement learning.

• Supervised learning-based intelligent sensing—Supervised learning deals with the known and labeled data and is divided into two types: classification and regression. This approach has been successfully implemented for many years in the fields of image classification, fraud detection, medical diagnosis, weather forecasting, market forecasting, and life expectancy estimation. In [52], ECG data are collected via wearable sensors, which detect heartbeats automatically, and a supervised learning approach is used for arrhythmia classification. An artificial haptic neuron system is fabricated in [53]. The system comprises a Nafion-based memristor and a piezoelectric sensor. The sensory receptor converts external stimulus into an electric signal, and the memristor is used for further processing of the data collected from the sensor. A supervised learning method is implemented for the recognition of English letters by placing the sensor on the joint of a finger. A novel methodology proposed in [54] using supervised learning for resolving the collision of cash tags yields high classification accuracy of listed companies in London Stock Exchange. A hybrid model that combines ML and game theory is proposed in [55] to solve issues related to network selection in ultra-dense heterogeneous networks.
  • K-Nearest Neighbors (K-NN) is an effective classification algorithm used for large datasets. Here, K represents the number of training samples that are near the test sample in the feature space [56]. In [57], a machine learning-based K-NN approach is used for load classification by collecting data from various smart plug sensors and other devices.
  • Support Vector Machine (SVM) is mainly used to categorize data attributes between classes by creating two-dimensional planes to minimize the classification error [58]. For example, Ref. [59] introduces a danger-pose detection system based on Wi-Fi devices that is used to monitor a bathroom while ensuring privacy. A machine learning-based detection approach usually requires large amount of data collected in target scenarios, which is challenging to detect danger situations. However, this work employed a machine learning-based anomaly-detection technique which requires a small amount of data in anomalous conditions. In this work, researchers first extracted the amplitude and phase shift from Wi-Fi Channel State Information (CSI) in order to detect low-frequency components associated with human activities. The static and dynamic features were then derived from the CSI changes over time. Finally, the static and dynamic characteristics are input into a one-class SVM which is employed as an anomaly-detection method to determine if a person is not in the bathtub, is bathing safely or in unsafe situations.
  • Decision Tree (DT) model consists of branches and nodes, wherein every node represents a test on every feature, and each branch has a value that the associated node can use to classify a sample [60]. A decision tree-based approach was presented in [61] for an intelligent transportation system (ITS). LIDAR sensors obtain point cloud data, which are then projected onto the XOY plane. After that, the images are classified into road and background grids for monitoring road traffic.
  • Ensemble Learning (EL) is a method based on combining the outputs of basic classification algorithms to boost the performance of classification. It is robust to data overfitting problem and is better than a single classifier [62]. This method is proposed in [63], where soft sensors are used to collect data to predict the composition, flow rate, and other features of the product, e.g., fatty acid methyl esters (FAME), in the procedure of production of biodiesel from vegetable oil.
  • Random Forest (RF) is made of a combination of several DTs and constructed randomly to form a model for improving the overall results [64]. A random forest-based classifier is proposed in [65] for estimating the content of bulky metals in agricultural soil using hyperspectral sensor data and is shown to reduce computational cost and time.
• Unsupervised learning-based intelligent sensing—Due to the large amount of unlabeled data in our everyday life, researchers have emphasized the unsupervised learning-based algorithms for intelligent sensing applications. This method consists of dimensionality reduction, generative networks, and clustering. Unsupervised learning-based intelligent sensing is proposed in [66], which is applied for real-time environment sensing to detect rare event instances intelligently. An unsupervised clustering-based method is introduced in [67] to describe an individual’s behavioral pattern by analyzing 100 days of unlabeled sensor data of 17 older adults from their homes and extract information of their day-to-day activities at different times. To detect the change in Landsat images, unsupervised learning is used in [68] with mean-shift clustering and hybrid wavelet transform under the Multi-Objective Particle Swarm optimization (MO-PSO) framework.
• Semi-Supervised intelligent sensing—This method deals with the combination of labeled and unlabeled data. To reduce the complexity of labeling all data for large datasets, semi-supervised methods are used. A robust model based on a semi-supervised approach is proposed in [69] to warn about the aircraft fault during the flight of a UAV by sensing real-time data such as angular velocity and pitch angle from flight sensors, and dramatically reduces the manual work. To detect faults in Additive Manufacturing (AM) products, a semi-supervised method with a few labeled data and a large number of unlabeled data is explored in [70].
• Reinforcement learning-based intelligent sensing—In the context of AI, reinforcement learning learns to make a sequence of decisions by interacting with its environment. One of the successful applications of this approach is to control autonomous cars by training the model. A deep reinforcement learning-based multi-sensor tracking fusion is proposed in [71] for vehicle tracking by learning on fused data from different sensors (camera and LIDAR). An intelligent sensing-based approach is introduced in [47] to autonomously monitor bridge conditions by collecting data from sensor nodes and make decisions using the reinforcement learning method. A novel approach based on YOLO V3 is proposed in [72] for multi-object tracking based on multi-agent deep reinforcement learning. This approach performs better in terms of precision, accuracy, and robustness. A routing protocol built on reinforcement learning is developed in [73] to find an optimal routing path for data transmission in a wireless network.

Table 3 shows the comparison of several ML and DL algorithms used in different areas of intelligent sensing. ML is a branch of AI that advocates the idea of acquiring the right data so that a machine can learn how to solve a particular problem by itself. The rise of ML is due to the availability of large datasets, and the adoption of ML algorithms in the field of intelligent sensing is to create smart devices that can take actions based on what they sense from the environment. With the implementation of ML in sensors, the efficiency and robustness of the system will reach the next level in smart sensing applications. Using sensor data, ML algorithms enable more robust predictions and classifications as compared to other physics-based models that envisage AI being added eventually to devices to adapt to the new circumstances. Therefore, the use of machine learning, including deep learning algorithms, is appropriate for performing challenging tasks in intelligent sensing, as shown in Figure 2.

The availability of datasets and the invention of new algorithms have increased the usage of ML and DL in the last few years. The supervised learning method has been used in numerous applications, such as object recognition, speech recognition, and spam detection. It predicts the value of one or more output variables (in the form of continuous or discrete) by observing input variables. The unsupervised learning method is generally used for gene clustering, social media analysis, and market research. The main focus of this method is to analyze unlabeled data. Semi-supervised learning is the hybrid model of supervised and unsupervised learning methods, which is used to solve problems with a few data points labeled and most of the data unlabeled. Reinforcement learning (RL) is used in applications such as finance, inventory management, and robotics, where the purpose is to learn a policy, i.e., to map situations between states of the environment to perform actions appropriately.

Deep Learning is now dominating the industry and research spheres for the growth of a range of smart-world systems for good reasons. DL has shown considerable potential in approximating and reducing huge datasets into accurate predictive and transformational output, greatly facilitating human-centered smart systems. This section discusses deep learning models based on intelligent sensing.

• Convolutional Neural Network—CNN is a robust supervised DL algorithm with better performance than other DL algorithms. IoT security is one of CNN’s applications where the features of the security data can be automatically learned by the sensors [82]. Deep CNN-based learning is proposed in [83] to recognize human emotions using electrodermal activity (EDA) sensors. These devices capture emotional patterns from a group of persons. The paper [84] proposed a system that detects the physical activity of older people from wearable sensors. For rotation-invariant features, each feature triplet is extracted from the X, Y, and Z axes and reduced to one feature represented by a 3D vector. Other works similar to this also achieve high accuracy in the study of younger people.
• Recurrent Neural Network—RNN is an important algorithm of DL in which present and past inputs depend on the output of the neural network. It is used to handle sequential inputs, which can be speech, text, or sensor data [85]. An RNN-based approach is discussed in [86], which is meant to interpolate sparse geomagnetic data from lost traces to reduce the time taken by linear interpolation approaches. The study in [87] discussed a mobile positioning method using RNN to analyze the strength of received signals. The authors experiment with the training of two RNNs separately for estimating latitude and longitude, which results in overfitting. An RNN-based learning model is proposed in [88] to monitor underwater sensor networks in real time, which improves the delay and reduces the cost of packet transmission.
• Generative Adversarial Network—GAN comprises two models; one is the generator, and the other is the discriminator. The two are trained in tandem via an adversarial process. These networks have been implemented for the security of IoT systems [89]. A conditional GAN-based DL method is presented for the reconstruction of CS-MRI that is compressed sensing magnetic resource imaging using compressed MR data [90]. In [91], the authors proposed a GAN-based method to generate X-ray prohibited images with different item poses. According to the paper, the quality of the images is good as compared to DC-GAN and WGAN-GP. After the images are generated, they are added to the real images and FID (Fréchet inception distance) is used to evaluate the performance of GANs.
• Long Short-Term Memory—LSTM is a type of recurrent neural network that is intended to model temporal sequences and their long-range dependencies more accurately than conventional RNNs [92]. The LSTM comprises units called memory blocks in the recurrent hidden layers. The memory blocks contain memory cells with self-connections that store the temporal state of the network in addition to special multiplicative units called gates to control the flow of information. A DL-based approach is used in [93] for emotion classification, dealing with a large number of sensor signals from different modalities. From the results presented in the paper, it came to be known that ad-hoc feature extraction may not be compulsory as DL models extract the high-level features automatically.

This section presents a review of some of the parameters that affect the performance of intelligent sensing. Intelligent sensing methods have been promising with state-of-the-art results in several areas, such as healthcare, image segmentation, agriculture, soft sensors, etc. The use of sensor systems in industrial, scientific, and consumer equipment is extensive and is continuously increasing in domains like automation. Essentially, industrial information revolutions need more sensors of every kind. The focus of the sensor system is to provide reliable signals and evaluate information. The smart sensing units include a sensing element and proper signal processing function within the same package.

Table 4 gives a list of parameters that affect the performance of intelligent sensing based on the results reported in literature. Key information includes the title and year of publication of each paper, and parameters that influence the performance of the various intelligent sensing approaches, such as temperature, accuracy, cost, time, occupancy, dependency, etc. One of the parameters is feature extraction in image recognition. Several techniques of pre-processing are used for enhancing certain features and removing unnecessary data. These techniques include digital spatial filtering, contrast enhancement, gray level distribution linearization, and image subtraction [94]. Measurement of redundancy in test samples is attempted to achieve test loss minimization, which can lead to a reduction of test maintenance costs and also ensure the integrity of test samples [95]. Evaluating ML algorithms is an important part of any project. Accuracy is one of the essential parameters to judge the performance of the trained model. Classification accuracy is defined as the fraction of correct predictions relative to the total number of input samples.

The most crucial aspect of this matter is the collection of data from multiple sources. The data usually goes through several stages of pre-processing to make it in presentable form. Intelligent sensing approaches are in general associated with technological applications where they are applied. For example, in cognitive radio, the sensing approach will be different from applications in a smart grid. The work in [102] presented an artificial intelligence-based approach for high-speed data delivery with latency regulation. Compared to CogMAC (Cognitive Medium Access Control) and AHP (Analytic Hierarchy Process) protocols, the decentralized approach helps in creating opportunistic methods for spectrum access and better design of channel selection mechanisms. The work presented in [103] proposed a method for integrating intelligence close to the sensor, which will enable decision-making in local nodes before transferring the information to the cloud or server. The local intelligence will be helpful in producing smart data that can be used for analysis to produce effective outcomes. Techniques such as normalization, linearization, and data cleaning can be done at local nodes in a piconet. Such inclusion will be helpful in the elimination of unnecessary steps, which needs to be done very frequently before data is used in artificial intelligence algorithms.

It is very important to identify data anomalies as data sometimes are collected from multiple platforms. In such cases, the source of data needs to be tracked for threat and irregularity. The work presented in [104] proposed scheduling and anomaly handling mechanisms in cross-platform IoT systems using cognitive tokens. The proposed methods use intelligent sensing with fair play and exponential growth procedures. In contrast to current technology trends in full-stack system development, a layered architecture-based approach was proposed in [105]. The proposed method will help to collect data, extract useful information, and transfer it for further processing. In the case of more sensitive data sensing, such as clinical or eHealth, Ref. [106] presented the implementation of gateway and scoring mechanisms to reduce the latency and to analyze the performance of systems. Such implementations have shown good performance in fog computing environments, where restricted resources are available at local nodes. The work presented in [107] shows the importance and challenges of IoT-based healthcare information sensing. The work presents challenges related to information acquisition, sensing, storage, processing, analytics, and presentation.